Brain inspired learning of dynamical systems

In this paper, we use dimensionality reduction techniques to study how a recurrent neural network (RNN) processes and encodes information in the context of a classification task, and we explain our findings using tools from dynamical systems theory. We observe that internal representations develop a task-relevant structure as soon as significant information is provided as input and this structure remains for some time even if we let the dynamics drift. However, the structure is only interpretable by the final classifying layer at the fixed time step for which the network was trained. We measure that throughout the training, the recurrent weights matrix is modified so that the resulting dynamical system associated with the network’s neural activations evolves into a non-trivial attractor, reminiscent of neural oscillations in the brain. Our findings suggest that RNNs change their internal dynamics throughout training so that information is stored in low-dimensional cycles, rather than in high-dimensional clusters.

Adaptive Games (AG) involve a controller agent that continuously feeds from player actions and game state to tweak a set of game parameters in order to maintain or achieve an objective function such as the flow measure defined by Csikszentmihalyi. This can be considered a Reinforcement Learning (RL) situation, so that classical Machine Learning (ML) approaches can be used. On the other hand, many games naturally exhibit an incremental gameplay where new actions and elements are introduced or removed progressively to enhance player’s learning curve or to introduce variety within the game. This makes the RL situation unusual because the controller agent input/output signature can change over the course of learning. In this paper, we get interested in this unusual “protean” learning situation (PL). In particular, we assess how the learner can rely on its past shapes and experience to keep improving among signature changes without needing to restart the learning from scratch on each change. We first develop a rigorous formalization of the PL problem. Then, we address the first elementary signature change: “input addition”, with Recurrent Neural Networks (RNNs) in an idealized PL situation. As a first result, we find that it is possible to benefit from prior learning in RNNs even if the past controller agent signature has less inputs. The use of PL in AG thus remains encouraged. Investigating output addition, input/output removal and translating these results to generic PL will be part of future works.

This thesis deals with recurrent neural networks, a particular class of artificial neural networks which can learn a generative model of input sequences. The input is mapped, through a feedback loop and a non-linear activation function, into a hidden state, which is then projected into the output space, obtaining either a probability distribution or the new input for the next time-step. This work consists mainly of two parts: a theoretical study for helping the understanding of recurrent neural networks framework, which is not yet deeply investigated, and their application to non-linear prediction problems, since recurrent neural networks are really powerful models suitable for solving several practical tasks in different fields. For what concerns the theoretical part, we analyse the weaknesses of state-of-the-art models and tackle them in order to improve the performance of a recurrent neural network. Firstly, we contribute in the understanding of the dynamical properties of a recurrent neural network, highlighting the close relation between the definition of stable limit cycles and the echo state property of an echo state network. We provide sufficient conditions for the convergence of the hidden state to a trajectory, which is uniquely determined by the input signal, independently of the initial states. This may help extend the memory of the network and increase the design options for the network. Moreover, we develop a novel approach to address the main problem in training recurrent neural networks, the so-called vanishing gradient problem. Our new method allows us to train a very simple recurrent neural network, making the gradient not to vanish even after many time-steps. Exploiting the singular value decomposition of the vanishing factors in the gradient and random matrices theory, we find that the singular values have to be confined in a narrow interval and derive conditions about their root mean square value. Then, we also improve the efficiency of the training of a recurrent neural network, defining a new method for speeding up this process. Thanks to a least square regularization, we can initialize the parameters of the network, in order to set them closer to the minimum and running fewer epochs of classical training algorithms. Moreover, it is also possible to completely train the network with our initialization method, running more iterations of it without losing in performance with respect to classical training algorithms. Finally, it is also possible to use it as a real-time learning algorithm, adjusting the parameters to the new data through one iteration of our initialization. In the last part of this thesis, we apply recurrent neural networks to non-linear prediction problems. We consider prediction of numerical sequences, estimating the following input choosing it from a probability distribution. We study an automatic text generation problem, where we need to predict the following character in order to compose words and sentences, and a path prediction of walking mobile users in the central area of a city, as a sequence of crossroads. Then, we analyse the prediction of video frames, discovering a wide range of applications related to the prediction of movements. We study the collision problem of bouncing balls, taking into account only the sequence of video frames without any knowledge about the physical characteristics of the problem, and the distribution over days of mobile user in a city and in a whole region. Finally, we address the state-of-the-art problem of missing data imputation, analysing the incomplete spectrogram of audio signals. We restore audio signals with missing time-frequency data, demonstrating via numerical experiments that a performance improvement can be achieved involving recurrent neural networks.

Concerning the learning problems of recurrent neural networks (RNNs), this paper deals with the problem of approximating a dynamical system (DS) by an RNN as one extension of the problem of approximating trajectories by an RNN. In particular, we systematically investigate how an RNN can produce a DS on the visible state space to approximate a target DS. First, it is proved that RNNs without hidden units uniquely produce a certain class of DSs. Next, a neural dynamical system (NDS) is proposed as such a DS that an RNN with hidden units can produce on the visible state space, and affine neural dynamical systems (A-NDSs) are constructed as concrete examples of NDSs. Moreover, we prove that any DS on a Euclidean space can be finitely approximated by some A-NDS with any precision, and propose adopting an A-NDS as such a DS that an RNN with hidden units produces to approximate a target DS. © 2000 Scripta Technica, Syst Comp Jpn, 31(4): 77–86, 2000

Learning and Synthesis Methods of Recurrent Spiking Neural Networks

Image segmentation refers to the process of dividing an image into multiple regions which represent meaningful areas. Image segmentation is an essential step for most image analysis tasks such as object recognition and tracking, pattern recognition, content-based image retrieval, etc. In recent years, a large number of image segmentation algorithms have been developed, but achieving accurate segmentation still remains a challenging task. Recently, reservoir computing (RC) has drawn much attention in machine learning as a new model of recurrent neural networks (RNN). Echo State Network (ESN) represents one efficient realization of RC, which is initially designed to facilitate learning in Recurrent Neural Networks. In this paper we investigate the viability of ESN as feature extractor for pixel classification based colour image segmentation. Extensive experiments are conducted on real world colour image datasets and the global ESN reservoir parameters are varied to identify their operating ranges that allow the use of the reservoir nodes internal activations as new pixel features for the colour image segmentation task. A simple feed forward neural network is used to realize the ESN readout function and classify these new features. The experimental results show that the proposed method achieves high performance image segmentation comparing with state-of-the-art techniques. In addition, a set of empirically derived guidelines for setting the reservoir global parameters are proposed.

Designing of VehiNet Using Convolutional Neural Networks and Deep Learning Techniques

To develop a natural language processing system that identifies relations of medications with adverse drug events from clinical narratives. This project is part of the 2018 n2c2 challenge. We developed a novel clinical named entity recognition method based on an recurrent convolutional neural network and compared it to a recurrent neural network implemented using the long-short term memory architecture, explored methods to integrate medical knowledge as embedding layers in neural networks, and investigated 3 machine learning models, including support vector machines, random forests and gradient boosting for relation classification. The performance of our system was evaluated using annotated data and scripts provided by the 2018 n2c2 organizers. Our system was among the top ranked. Our best model submitted during this challenge (based on recurrent neural networks and support vector machines) achieved lenient F1 scores of 0.9287 for concept extraction (ranked third), 0.9459 for relation classification (ranked fourth), and 0.8778 for the end-to-end relation extraction (ranked second). We developed a novel named entity recognition model based on a recurrent convolutional neural network and further investigated gradient boosting for relation classification. The new methods improved the lenient F1 scores of the 3 subtasks to 0.9292, 0.9633, and 0.8880, respectively, which are comparable to the best performance reported in this challenge. This study demonstrated the feasibility of using machine learning methods to extract the relations of medications with adverse drug events from clinical narratives.

Recurrent neural networks (RNN) are ubiquitous computing systems for sequences and multivariate time-series data. While several robust RNN architectures are known, it is unclear how to relate RNN initialization, architecture, and other hyperparameters with accuracy for a given task. In this work, we propose treating RNN as dynamical systems and correlating hyperparameters with accuracy through Lyapunov spectral analysis, a methodology designed explicitly for nonlinear dynamical systems. To address the fact that RNN features go beyond the existing Lyapunov spectral analysis, we propose to infer relevant features from the Lyapunov spectrum with an Autoencoder and an embedding of its Latent representation (AeLLE). Our studies of various RNN architectures show that AeLLE successfully correlates RNN Lyapunov spectrum with accuracy. Furthermore, the Latent representation learned by AeLLE is generalizable to novel inputs from the same task and is formed early in the process of RNN training. The latter property allows for predicting the accuracy to which RNN would converge when training is complete. We conclude that the representation of RNN through the Lyapunov spectrum, along with AeLLE, provides a novel method for the organization and interpretation of variants of RNN architectures.

Protein secondary structure prediction using modular reciprocal bidirectional recurrent neural networks

Cortical networks are capable of unsupervised learning and spontaneous replay of complex temporal sequences. Endowing artificial spiking neural networks with similar learning abilities remains a challenge. In particular, it is unresolved how different plasticity rules can contribute to both learning and the maintenance of network stability during learning. Here we introduce a biologically inspired form of GABA-Modulated Spike Timing-Dependent Plasticity (GMS) and demonstrate its ability to permit stable learning of complex temporal sequences including natural language in recurrent spiking neural networks. Motivated by biological findings, GMS utilizes the momentary level of inhibition onto excitatory cells to adjust both the magnitude and sign of Spike Timing-Dependent Plasticity (STDP) of connections between excitatory cells. In particular, high levels of inhibition in the network cause depression of excitatory-to-excitatory connections. We demonstrate the effectiveness of this mechanism during several sequence learning experiments with character- and token-based text inputs as well as visual input sequences. We show that GMS maintains stability during learning and spontaneous replay and permits the network to form a clustered hierarchical representation of its input sequences. Overall, we provide a biologically inspired model of unsupervised learning of complex sequences in recurrent spiking neural networks.

Aims This research paper aims to check the effectiveness of a variety of machine learning models in classifying esophageal cancer through MRI scans. The current study encompasses Convolutional Neural Network (CNN), K-Nearest Neighbor (KNN), Recurrent Neural Network (RNN), and Visual Geometry Group 16 (VGG16), among others which are elaborated in this paper. This paper aims to identify the most accurate model to facilitate increased, improved diagnostic accuracy to revolutionize early detection methods for this dreadful disease. The ultimate goal is, therefore, to improve the clinical practice performance and its results with advanced machine learning techniques in medical diagnosis. Background Esophageal cancer poses a critical problem for medical oncologists since its pathology is quite complex, and the death rate is exceptionally high. Proper early detection is essential for effective treatment and improved survival. The results are positive, but the conventional diagnostic methods are not sensitive and have low specificity. Recent progress in machine learning methods brings a new possibility to high sensitivity and specificity in the diagnosis. This paper explores the potentiality of different machine-learning models in classifying esophageal cancer through MRI scans to complement the constraints of the traditional diagnostics approach. Objective This study is aimed at verifying whether CNN, KNN, RNN, and VGG16, amongst other advanced machine learning models, are effective in correctly classifying esophageal cancer from MRI scans. This review aims at establishing the diagnostic accuracy of all these models, with the best among all. It plays a role in developing early detection mechanisms that increase patient outcome confidence in the clinical setting. Methods This study applies the approach of comparative analysis by using four unique machine learning models to classify esophageal cancer from MRI scans. This was made possible through the intensive training and validation of the model using a standardized set of MRI data. The model’s effectiveness was assessed using performance evaluation metrics, which included accuracy, precision, recall, and F1 score. Results In classifying esophageal cancers from MRI scans, the current study found VGG16 to be an adequate model, with a high accuracy of 96.66%. CNN took the second position, with an accuracy of 94.5%, showing efficient results for spatial pattern recognition. The model of KNN and RNN also showed commendable performance, with accuracies of 91.44% and 88.97%, respectively, portraying their strengths in proximity-based learning and handling sequential data. These findings underline the potential to add significant value to the processes of esophageal cancer diagnosis using machine learning models. Conclusion The study concluded that machine learning techniques, mainly VGG16 and CNN, had a high potential for escalated diagnostic precision in classifying esophageal cancer from MRI imaging. VGG16 showed great accuracy, while CNN displayed advanced spatial detection, followed by KNN and RNN. Thus, the results set new opportunities for introducing advanced computational models to the clinics, which might transform strategies for early detection to improve patient-centered outcomes in oncology.

Electromyography (EMG) signals provide vital insights into muscular dynamics, making them indispensable for gesture recognition tasks. However, conventional deep learning models often suffer from high power consumption, limiting their deployment on neuromorphic edge hardware. This study proposes recurrent spiking neural networks (RSNNs) based framework for EMG classification with multiple neuron configurations involving Integrate and Fire (IF) and Leaky Integrate and Fire (LIF) neuron models. These configurations were analyzed to result in better trade-off between temporal stability and computational efficiency. Among various neuron cluster configurations within the reservoir, entirely IF neuron based network achieved the highest classification accuracy of 80%. This also outperforms existing deep learning based approaches. These results underscore the potential of RSNNs as an efficient framework for real-time EMG gesture recognition in neuromorphic hardware.

Use of deep learning for sequence dependent task classification has shown considerable improvement in recent past. This paper is on task identification from the associated electroencephalograph (EEG) signal. This paper deals with extraction of EEG sub-band features by using wavelet transform followed through classification by means of recurrent neural network (RNN) and deep learning. The methodology is based on comparison of correlation among various EEG spectral components by using a holdup RNN architecture and winner-takes-all logistic regression approach to identify the associated task. The bottleneck of current EEG processing approaches is signal power getting biased due to noisy trials which deteriorates the analysis in the manifold space. This study uses a similarity or coherence analysis between sub-band spectrum of individual trial which invalidates the effect of noise. To assess the proposed method, a dataset, containing 5 subjects performing four mental tasks, is utilized. The results show that in the multiclass scenario the proposed algorithm performs at par with state-of-art results, whereas in two class scenario, the proposed algorithm outperforms most of the reported results.

Recurrent neural networks (RNNs) have been successfully applied to a variety of problems involving sequential data, but their optimization is sensitive to parameter initialization, architecture, and optimizer hyperparameters. Considering RNNs as dynamical systems, a natural way to capture stability, i.e., the growth and decay over long iterates, are the Lyapunov Exponents (LEs), which form the Lyapunov spectrum. The LEs have a bearing on stability of RNN training dynamics since forward propagation of information is related to the backward propagation of error gradients. LEs measure the asymptotic rates of expansion and contraction of non-linear system trajectories, and generalize stability analysis to the time-varying attractors structuring the non-autonomous dynamics of data-driven RNNs. As a tool to understand and exploit stability of training dynamics, the Lyapunov spectrum fills an existing gap between prescriptive mathematical approaches of limited scope and computationally-expensive empirical approaches. To leverage this tool, we implement an efficient way to compute LEs for RNNs during training, discuss the aspects specific to standard RNN architectures driven by typical sequential datasets, and show that the Lyapunov spectrum can serve as a robust readout of training stability across hyperparameters. With this exposition-oriented contribution, we hope to draw attention to this under-studied, but theoretically grounded tool for understanding training stability in RNNs.